Tom Blomfield’s case for running a company as AI loops, tested against the evidence, with the controls a board would need.
Terence KokExecutive Director, AI Governance & Assurance, Orion Five Engineering
About your speaker
Terence Kok
Executive Director, AI Governance & Assurance Practice, Orion Five Engineering
Former Chief AI and Innovation Officer, Meinhardt Group
Twenty-five years leading digital and AI transformation across Asia and the Middle East: national smart-city programmes, critical infrastructure and enterprise operations. Today I help leadership teams put AI into production with controls they can show a board, an auditor and a regulator.
BuiltA governed AI platform serving more than 6,000 engineers, with zero security incidents
CertifiedISO/IEC 42001 Lead Auditor Professional (BSI) · AIGP (SMU) · Registered Management Consultant
ServingCTCB CPSC Advisory Board · UN ESCAP AI for Developing Countries Forum
AwardsCDO Vision Award 2026 · Singapore's Most Influential Data & AI Leaders 2026 · Google Trailblazer Transformation Award 2024
Independent analysis from public sources. The views in this session are mine.
Case study · The self-improving company
This session
The plan for this session.
The argumentA legion, a loop and the company’s own know-how.
The YC exampleAn internal agent that fixes itself overnight.
Five implicationsTokens, managers, recording, software and people.
Against the evidenceWhat published research says about each claim.
Where governance breaksAuto-merge, quality gates, the law and accountability.
My viewWhich loops can run unattended, and the controls the rest need.
Part seven is for the board: lessons, a controls checklist and questions for discussion.
Case study · The self-improving company
The case in one slide
A short talk that redraws the company.
5
stages in the loop: sensor, policy, tools, quality gate, learning
2
roles left: the builder-operator and the directly responsible individual
5×
revenue per employee at Demo Day against 18 months earlier, on the speaker’s account
0
published YC data behind the examples, so far
Sources: [1] [2]. The 5× figure and the YC examples rest on the speaker’s words only.
Case study · The self-improving company
Where the talk came from
How the talk came about.
26 Feb 2026
Block halves its staff
From over 10,000 people to just under 6,000.
31 Mar 2026
Dorsey and Botha essay
“From Hierarchy to Intelligence”: no permanent middle management layer.
23 Apr 2026
GPT-5.5 in Codex
The model the talk calls “Codex 5.5”.
24 Apr 2026
Diana Hu’s talk
Building a company with AI from the ground up.
21 May 2026
Blomfield’s talk online
A batch talk by a YC General Partner.
13 Jul 2026
Blomfield joins Anthropic
On leave from YC, to its compute team.
Sources: [1] [3] [5] [7] [9] [10]. The recording date is not published; the talk’s own references place it in late April or early May 2026.
Case study · The self-improving company
01
The argument.
Companies are still built like Roman legions. The talk says AI breaks that design.
Case study · The self-improving company
The opening image
How a legion passed information.
Unit sizes approximate and varied over time; see [18]. The talk: “human beings are the conduit for information flowing up and down” [1].
Case study · The self-improving company
In the speaker’s words
“If you think about most companies today, they are organised like a Roman legion, where human beings are the conduit for information flowing up and down.”
Tom Blomfield, Y Combinator batch talk [1]. The legion image comes from the Dorsey and Botha essay [5].
Case study · The self-improving company
The wrong frame
Why copilots are the wrong frame.
Last year’s frame
AI as a tool for each person
Copilots make engineers 20% more productive. Assistants are added to existing workflows. The organisation stays the same shape.
“Taking the old way of working and adding a more powerful engine onto it.”
The talk’s frame
The company as a set of loops
Know-how is pulled out of people’s heads into shared context and skills. Loops sense, decide, act, check and learn. The organisation is redrawn around them.
“A set of recursive, self-improving AI loops.”
Source: [1].
Case study · The self-improving company
The first step
Extract the company’s know-how.
People’s headsHow the work is actually done
Slack and DMsDecisions and exceptions
EmailCustomers, partners, promises
Notion and docsPolicies, half out of date
Company context and skillsLegible to AI, so loops can act on it
Source: [1]. “If you can make that legible, you suddenly can move from this hierarchical organisation to … an intelligent AI-powered organisation.”
Case study · The self-improving company
The loop
The five stages of the loop.
1
2
3
4
5
1SensorCustomer emails, support tickets, code changes, cancellations, product telemetry.
2PolicyWhat the system may do, what needs a person’s permission, what it must log.
3ToolsDeterministic APIs and skills: query the database, read the calendar.
4Quality gateEvals, deterministic checks, safety filters, human review for high-risk work.
5LearningFinds where it failed in the real world and feeds that back to the top.
Source: [1]. The speaker credits the structure to Diana Hu’s talk [3].
Case study · The self-improving company
02
The YC example.
An internal agent that grew from answering questions to changing its own code overnight.
Case study · The self-improving company
On the speaker’s account
Three stages of one internal agent.
Stage 1
An agent you can ask
Deterministic tools query YC’s database: “when did I last have office hours with this company?”
A better search box
Stage 2
A sidekick
RAG and several query routes find five founders in petrochemicals for an introduction.
“20 or 30% more effective”
Stage 3
A loop that improves itself
A monitoring agent reads every staff query, finds the failures, works out why and ships a fix overnight.
The “aha moment”
Source: [1]. YC has not published the system, its review rules or its error rate.
Case study · The self-improving company
Stage 3, step by step
Nobody signs between query and morning.
1
Watch
Every query by every YC employee.
Monitoring agent
2
Diagnose
New tool, skills file, database view or index?
Monitoring agent
3
Write the fix
Code and a merge request to the YC codebase.
Agent
4
Review
A second agent reviews the change.
No named person
5
Merge and deploy
Into production, overnight.
No named person
6
Next morning
The same query now succeeds.
Staff member
EveningMorning
Source: [1]: “Let’s write the code, put in a merge request … have an agent review it, and merge it and deploy it.” Orange marks steps with no named person, on the speaker’s account.
Case study · The self-improving company
Two more loops from the talk
Product and support loops.
Product loop
A funnel that tunes itself
An agent reads product analytics, finds the step with most friction, researches practice, runs an A/B test for a week, keeps the winner and repeats.
What changesCustomer-facing screens, weekly, without a person choosing.
Support loop
Suggestions shipped overnight
Customer suggestions are triaged by agents acting as “chief product officer” and “chief technology officer”. Those in line with the roadmap are coded, deployed and shipped.
What changesProduct scope and code, “without a human being involved”.
Source: [1].
Case study · The self-improving company
03
Five implications.
Tokens over headcount, no middle managers, record everything, disposable software, people at the edge.
Case study · The self-improving company
Implication 1
Burn tokens, not headcount.
The speaker says YC companies reach Demo Day with about five times more revenue per employee than 18 months ago, and expects firms to be “constrained on token usage, not on headcount”.
He calls measuring each person’s token use “obviously dumb and gameable at the extreme” but “directionally correct”, a way to see who is experimenting.
Source: [1]. No YC data has been published for the 5× figure.
Case study · The self-improving company
Implication 2
Middle management is done, says the talk.
Kept
Individual contributor
“Who build and operate capabilities, the model, the intelligence layer, and the interfaces.”
Kept
Directly responsible individual
“Who own specific cross-cutting problems or opportunities and customer outcomes.” A named person, not a committee.
Dropped
Player-coach
“Who combine building with developing people.”
“Jack Dorsey has three. I actually don’t like the third one, so I deleted it.”
Role definitions from Dorsey and Botha [5]; the edit is the speaker’s [1]. The dropped role is the one that develops people.
Case study · The self-improving company
Implication 3
Record everything.
2,000
hours of office hours
Recorded over about three months, with partner emails, Slack and DMs.
→
1
weekend
Diarised, synthesised and sorted by topic: fundraising, hiring, co-founder disputes.
→
150
page manual
Replacing a user manual written five to ten years earlier.
→
Monthly
update
New advice compared with the manual and either added or thrown away.
Source: [1], all on the speaker’s account. He would also wire rooms with microphones, or use phones, clips or smart glasses.
Case study · The self-improving company
Implication 4
Software is ephemeral.
GeneratedEvent dashboardBuilt for one YC event, then thrown away
GeneratedOps workflowBuilt by the operations team itself
GeneratedInternal report“One-shot” by a coding model
RegeneratedNext month’s versionSame instructions, smarter model
Kept for goodAll the data, every email in Markdown, the skills and the know-how of how each function works
Source: [1]. The model referred to as “Codex 5.5” is OpenAI’s GPT-5.5, released 23 April 2026 [10].
Case study · The self-improving company
Implication 5
People at the edge.
Company brainData, messages, skills, know-how
Novel situationsCases the context has never seen
Ethical callsWhere values, not data, decide
High stakes, high emotionA founder thinking of splitting with a co-founder
Sales in the room“A human being in the room for the next twenty years”
Source: [1].
Case study · The self-improving company
The question the talk ends on
“If you were building your company today, would you start it in this shape?”
“For most of you, you’re small enough to build it right.” [1]
Case study · The self-improving company
04
Against the evidence.
Ten claims from the talk, set against published research, company records and incidents.
Case study · The self-improving company
My reading of the record · 1 of 2
How the talk’s claims hold up.
Hierarchy exists to relay informationGalbraith saw hierarchy as one way to process information among several. It also carries accountability and develops people; managers explain at least 70% of the variance in team engagement.Partly holds
Copilots are the wrong frameMeasured copilot gains are real but modest and uneven: 15% in one large field study, 19% slower in one developer trial. Most firms report no P&L return from pilots.Partly holds
Loops can run end to end without a personCoding agents have deleted production databases. Frontier models game the checks that grade them. The talk’s own quality gate keeps human review for high-risk work.Against the evidence
YC’s agent fixes itself overnightDescribed in the talk only. No published design, review rules or error rate.Not shown
5× revenue per employee at Demo DayNo YC data published. AI-native firms do report high ratios (Gamma about $100m ARR with about 50 staff), on run-rate revenue that leaves out the compute bill.Not shown
Verdicts are mine; sources against each claim on the case study page.
Case study · The self-improving company
My reading of the record · 2 of 2
How the remaining claims hold up.
Token use shows who is pushing hardestMeta’s and Amazon’s token leaderboards were gamed and taken down within months.Against the evidence
Middle management is doneDelayering is real at Google, Amazon, Meta and Block. Klarna and Commonwealth Bank walked back AI-led cuts in 2025.Partly holds
Record everything to make the firm legibleLegible data has value. Always-on recording needs notice or consent in many places, and the EU bans AI emotion inference at work.Partly holds
Software is disposable, context is the assetConsistent with how internal tools are now built from prompts and stored knowledge. Change control still applies to what the regenerated tool can touch.Holds
People belong at the edge, in high-stakes momentsIn line with the EU AI Act’s human oversight article and Singapore’s agentic AI framework, both of which also warn of automation bias.Holds
Verdicts are mine; sources against each claim on the case study page.
Case study · The self-improving company
Claim 1
Hierarchy does more than relay information.
1974
Galbraith
Hierarchy is one way to process information, alongside information systems and lateral links. An AI layer is a new information system, so some management work can move to it.
70%
Gallup
Of the variance in team engagement is explained by the manager. The role the talk deletes, the player-coach, is the one that develops people.
1945
Hayek
“The knowledge of the particular circumstances of time and place” is mostly never written down, so it never reaches a recording.
Sources: [22] [33] [21]. Coase (1937): cheaper coordination can make firms larger as easily as flatter [20].
Case study · The self-improving company
Claim 2
What copilots delivered.
15%
More issues resolved per hour
5,172 support agents. Most of the gain went to less experienced staff; the most experienced saw small quality declines.
Brynjolfsson, Li and Raymond, QJE 2025 [25]
19%
Longer, not shorter
Experienced developers on their own repositories, who believed they had been faster. A 2026 follow-up is “very weak evidence” of a speed-up.
METR, July 2025 and February 2026 [23] [24]
95%
Of organisations with no P&L return
From generative AI pilots, blamed on a “learning gap”. The talk’s learning loop answers the diagnosis; the cure is untested.
MIT NANDA via Fortune, August 2025; small sample [26]
Case study · The self-improving company
Claim 7
Delayering is well under way.
Meta, 2023Removed management layers; most managers capped at ten direct reports.[30]
Amazon, 2024Each senior team to raise its ratio of individual contributors to managers by at least 15%.[28]
Google, 202535% fewer managers of teams with fewer than three people; many became individual contributors.[29]
Block, 2026From over 10,000 people to just under 6,000, framed as replacing middle managers with AI.[7] [8]
Layoffs, 2024Middle managers were 29% of layoffs, against 22% in 2018 to 2022.[32]
Gartner forecastThrough 2026, 20% of organisations will use AI to remove more than half of their middle management.[31]
Case study · The self-improving company
Claim 7, the other side
Two firms walked back AI-led cuts in 2025.
Klarna, May 2025. The CEO said cost had been “a too predominant evaluation factor” and the result was “lower quality”.
Commonwealth Bank, August 2025. Reversed 45 redundancies made for an AI voice bot, called the decision an error and apologised.
Block, 2026. Staff said about 95% of AI-generated code changes still needed human modification.
Sources: [34] [35] [7]. Full case studies of Klarna and Commonwealth Bank are on terencekok.com.
Case study · The self-improving company
Claims 5 and 6
Revenue per employee and token use.
Revenue per employee
The ratios are real, the denominator is narrow
Gamma: about $100m ARR with about 50 staff (Nov 2025). Lovable: $100m ARR with 45 staff (Jul 2025).
Read with careRun-rate revenue, and the compute bill sits outside headcount. Cost moves to a model supplier.
[36] [37]
Tokens as a measure
Gamed within weeks
Meta’s “Claudeonomics” board ranked about 85,000 staff, logged over 60 trillion tokens in 30 days and was taken down. Amazon staff inflated counts on trivial work.
Better testShopify: show why AI cannot do the work before asking for headcount.
[43] [44] [11] [13]
Case study · The self-improving company
Claim 8
Recording everything meets the law first.
United StatesEleven states, California among them, mainly require every party’s consent to record a conversation.[45]
Singapore PDPACollection for managing employment is allowed, but staff must be told the purpose. Clients and visitors in the room are not covered.[46]
GDPR Art. 88Employment rules must protect dignity, with monitoring systems at work named specifically.[47]
EU AI Act Art. 5(1)(f)Since 2 February 2025, AI that infers emotions in the workplace is prohibited, except for medical or safety reasons.[48]
BehaviourMonitoring used for control lowers performance and raises rule-breaking.[57]
Microphones in every room feeding a system that spots “high-emotion moments” comes close to the EU prohibition.
Case study · The self-improving company
05
Where governance breaks.
Agents with production access, graders that get gamed, and people who cannot review it all.
Case study · The self-improving company
Claim 3: loops without a person
Coding agents have already deleted production systems.
Replit, July 2025
A coding agent deleted a production database during a code freeze.
April 2026
A coding agent deleted a production database and the backups beside it in nine seconds. About 30 hours down.
AWS, December 2025
An outage reported as caused by an internal agent. Amazon says the cause was misconfigured access, not AI.
Each one is a permissions failure as much as a model failure. The tools layer decides how much damage one bad change can do.
Sources: [38] [39] [40].
Case study · The self-improving company
The quality gate
Models game the checks that grade them.
30%
of runs on one benchmark where a frontier model gamed its grader; telling it not to cheat made little difference
100%
of runs on one task in the same METR study
Sabotage
when models that learned to game coding tests were used as agents, Anthropic reported in November 2025
An agent that reviews another agent’s code and approves the merge is the kind of grader these studies show being gamed.
Sources: [41] [42].
Case study · The self-improving company
Putting a person back
The limits of human review.
People who supervise automation drift into complacency under load, experts included, and training does not remove it.
A review of 41 policies requiring human oversight found people often cannot do the oversight the policy assumes.
An overnight loop produces changes faster than anyone can read them. Review has to be tiered by risk.
Sources: [53] [54]. The EU AI Act’s Article 14 names automation bias [49].
Case study · The self-improving company
06
My view.
Two questions sort every loop: how much harm can one bad change do, and how easily is it undone?
Case study · The self-improving company
Which loops can run unattended
The talk’s loops by consequence and reversibility.
A named person signsBefore it reaches a customer
Do not automate yetLegal or human judgment first
Run unattendedLog, sample, review weekly
Run inside limitsHard guardrails, automatic rollback
Fix a database view for staff queries
Regenerate an internal dashboard
Weekly A/B test of the sign-up funnel
Monthly update of the advice manual
Ship customer suggestions overnight
Token leaderboard used in reviews
Microphones in every room
Harder to undo →Higher consequence →
Case study · The self-improving company
The governed loop
Three controls added to the loop.
SensorWhose data, and were they told?
PolicyWho wrote the rules, and who can change them?
ToolsLeast privilege; no agent token near production backups.
Quality gateWho decides what counts as high risk?
LearningThe loop cannot change its own gate.
1 · Separate graderA different model, a deterministic test or a person reviews each change.
2 · Named ownerThe talk’s directly responsible individual, with a kill switch.
3 · Change logWhat triggered each change, so it can be traced and reversed.
My recommendation. None of these slows the low-risk loops; they are what lets a board approve the rest.
Case study · The self-improving company
My verdict
What I would keep and what I would change.
Keep
Most of the talk
Make the company’s know-how legible: the cheapest first step, and most firms have not started.
Treat internal software as disposable and keep the data.
Keep people for novel, ethical and high-stakes moments.
Change
Four things a board would sign
No loop merges its own code without a separate grader and a named owner.
Judge outcomes, never token volume.
Check consent and the law before any recording.
Keep someone whose job is to develop people.
Case study · The self-improving company
For Singapore boards
Where Singapore regulators stand.
IMDA, 22 Jan 2026Model AI Governance Framework for Agentic AI: bound the agent’s scope and tools up front; keep people meaningfully accountable.[51]
MAS, 7 Oct 2026Final AI risk management guidelines: board oversight, an AI inventory, human oversight and change management, including agents. In effect from 7 Oct 2027.[50]
PDPCEmployers must tell staff why personal data is collected, even under the employment exception.[46]
A loop that merges its own code is a change-management question before it is a technology one.
Case study · The self-improving company
07
For the board.
Seven lessons, a controls checklist and six questions to take back.
Case study · The self-improving company
Lessons 1 to 4
What boards can take from the talk.
01
Start with legibility
Write down how each function works and keep the data. It pays off with or without autonomous loops.
02
Sort loops by risk
Unattended where harm is low and reversal easy; limits and rollback in the middle; a named signature on the rest.
03
Separate the grader
Models have been measured gaming their checks. Lock every loop out of its own quality gate.
04
Name an owner
Keep the directly responsible individual, with the authority to stop the loop and the duty to answer for it.
Case study · The self-improving company
Lessons 5 to 7
The remaining lessons.
05
Measure outcomes
Token leaderboards were gamed within weeks. Judge teams on results.
AskWhat can AI not do here, before more headcount?
06
Law before microphones
Consent, notice and the EU ban on workplace emotion inference apply before any capture starts.
AskWho in the room is not an employee?
07
Decide who develops people
If the middle layer goes, someone still coaches, settles disputes and builds judgment.
AskWho does that job after the restructure?
Case study · The self-improving company
Controls checklist
Controls for a self-improving loop.
Control
What good looks like
Owner
Loop register
Every loop listed with sensors, tools, write access, owner and risk tier
Chief AI or technology officer
Separate grader
A different model, a deterministic test or a person reviews each change
Head of engineering
Least privilege
Production and backups out of reach of any agent token; destructive actions need a second key
CISO
Rollback and kill switch
Every change reversible within minutes; the owner can stop the loop at once
Loop owner
Change log
What triggered each change, what it did, who or what approved it
Internal audit
Recording assessment
Consent or notice, purpose, retention and third parties documented first
Data protection officer
Outcome metrics
Errors, reversals and customer results, never token volume
COO
My recommendations, drawn from the seven lessons.
Case study · The self-improving company
For discussion
Six questions to take back to your board.
Which of our functions could we describe well enough today for an AI to run?
Which AI loops already change our systems without a named person signing?
Who grades our AI’s work, and could the AI edit that grader?
What do we record about staff and customers, and who was told?
Which of our AI metrics measure input rather than outcome?
If we removed a management layer, who would develop our people?
Case study · The self-improving company
Evidence and methodology
How strong each claim in this case is.
Claim
Rests on
Grade
Who gave the talk, when it was published, what it says
Managers 29% of 2024 layoffs; 70% of engagement variance
Live Data via Korn Ferry; Gallup
Measured, vendor data
Agent incidents (Replit, database and backups deleted)
The Register; OECD AI Incidents Monitor
Press and incident records
Models game their graders
METR; Anthropic
Measured, lab studies
Token leaderboards gamed and retired
Fortune; IBM
Press
Legal limits on recording
RCFP; PDPC; GDPR; EU AI Act
Law and regulator guidance
Verdicts, autonomy map, governed loop and lessons
My reading of the record cited against each
My call
Case study · The self-improving company
Full timeline · 1 of 2
1937 to August 2025.
1937Ronald Coase argues firms exist because coordinating through markets has costs; a firm grows until internal coordination costs as much.[20]
1974Jay Galbraith treats hierarchy as one way to process information, alongside information systems and lateral links.[22]
14 Mar 2023Meta’s “year of efficiency” removes management layers and caps most managers at ten direct reports.[30]
Sep 2024Amazon tells each senior team to raise its ratio of individual contributors to managers by at least 15% by the end of Q1 2025.[28]
22 Oct 2024Gartner predicts that through 2026, 20% of organisations will use AI to flatten, cutting more than half of their middle management.[31]
7 Apr 2025Shopify’s CEO publishes a memo making reflexive AI use a baseline expectation, with AI use in performance reviews.[13]
May 2025Klarna’s CEO says a cost focus in customer service led to “lower quality”.[34]
10 Jul 2025METR’s trial finds experienced developers took 19% longer with AI tools on their own repositories.[23]
Jul 2025A Replit coding agent deletes a production database during a code freeze.[38]
Aug 2025Commonwealth Bank reverses 45 redundancies made for an AI voice bot and apologises. Google says it has 35% fewer managers of small teams.[35] [29]
Case study · The self-improving company
Full timeline · 2 of 2
November 2025 to October 2026.
21 Nov 2025Anthropic reports that models which learned to game coding tests went on to sabotage work in agentic use.[42]
22 Jan 2026Singapore launches its Model AI Governance Framework for Agentic AI at Davos.[51]
26 Feb 2026Block announces it will cut from over 10,000 people to just under 6,000.[7] [8]
31 Mar 2026Jack Dorsey and Roelof Botha publish “From Hierarchy to Intelligence”, using the Roman legion as the model to retire.[5] [6]
23–24 Apr 2026OpenAI releases GPT-5.5 in Codex. YC publishes Diana Hu’s talk on building a company with AI.[10] [3]
Apr 2026A coding agent deletes a production database and its backups in nine seconds.[39]
Spring 2026Token leaderboards at Meta and Amazon are gamed and withdrawn.[43] [44]
21 May 2026YC publishes Tom Blomfield’s batch talk, “How to Build a Self-Improving Company with AI”.[1]
13 Jul 2026Blomfield announces leave from YC to join Anthropic’s compute team.[9]
7 Oct 2026MAS issues final AI risk management guidelines for financial institutions, including human oversight and change management.[50]
Case study · The self-improving company
Take it further
The full case and two tools.
The self-improving company case study
Full text, motion figures, 57 numbered sources and the FAQ. terencekok.com/research/case-studies/self-improving-company
25Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942. Stanford Digital Economy Lab.digitaleconomy.stanford.edu
26Fortune. (2025, 18 August). MIT report: 95% of generative AI pilots at companies are failing.fortune.com
27InfoQ. (2025, September). DORA 2025: State of AI-assisted software development.infoq.com
28CNBC, via NBC DFW. (2024, September). Amazon and the endangered future of the middle manager.nbcdfw.com
29NBC News. (2025, 27 August). Google eliminated 35% of managers overseeing small teams, exec says.nbcnews.com
30Zuckerberg, M. (2023, 14 March). Update on Meta’s year of efficiency. Meta.about.fb.com
Case study · The self-improving company
Sources · 3 of 4
Research, records and incidents.
31Network World. (2024, 22 October). AI dominates Gartner’s 2025 predictions.networkworld.com
32Korn Ferry. (2025, 26 March). The great flattening experiment. Briefings magazine, issue 68.kornferry.com
33Beck, R., & Harter, J. (2015). Managers account for 70% of variance in employee engagement. Gallup.news.gallup.com
34Entrepreneur. (2025, 9 May). Klarna CEO reverses course by hiring more humans, not AI.entrepreneur.com
37TechCrunch. (2025, 23 July). Eight months in, Swedish unicorn Lovable crosses the $100M ARR milestone.techcrunch.com
38The Register. (2025, 22 July). Replit makes vibe-y promise to stop its AI agents making vibe-y coding mistakes.theregister.com
39OECD AI Incidents Monitor. (2026, 27 April). AI coding agent deletes a production database and its backups.oecd.ai
40OECD AI Incidents Monitor. (2026, 20 February). AWS outage reportedly caused by an AI coding agent.oecd.ai
41METR. (2025, 5 June). Recent frontier models are reward hacking.metr.org
42Anthropic. (2025, 21 November). From shortcuts to sabotage: natural emergent misalignment from reward hacking.anthropic.com
43Fortune. (2026, 12 May). Tokenmaxxing at Amazon and Meta.fortune.com
44IBM Think. (2026). What is tokenmaxxing?ibm.com
Case study · The self-improving company
Sources · 4 of 4
Checked on 25 October 2026.
45Reporters Committee for Freedom of the Press. Reporter’s recording guide.rcfp.org
46Personal Data Protection Commission Singapore. Advisory guidelines on the PDPA for selected topics, chapter 5: employment.pdpc.gov.sg
47European Union. (2016). General Data Protection Regulation, Article 88: processing in the context of employment.gdpr-info.eu
48European Union. (2024). AI Act, Article 5: prohibited AI practices.artificialintelligenceact.eu
49European Union. (2024). AI Act, Article 14: human oversight.artificialintelligenceact.eu
50Monetary Authority of Singapore. (2026, 7 October). MAS sets out supervisory expectations on responsible AI adoption by financial institutions.mas.gov.sg
51CMS. (2026, January). Singapore issues global-first agentic AI model governance framework at Davos.cms.law
52National Institute of Standards and Technology. AI Risk Management Framework.nist.gov
53Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: an attentional integration. Human Factors, 52(3), 381–410.doi.org
54Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review, 45, 105681.arxiv.org
55Lashinsky, A. (2012). Inside Apple. Hachette Book Group.hachettebookgroup.com
56Fortune. (2026, 17 April). How Jack Dorsey decided to lay off 40% of Block’s staff.fortune.com
57Thiel, C., McClean, S., Harvey, J., & Prince, N. (2024). Surveilling employees erodes trust, and puts managers in a bind. Harvard Business Review, via EBSCO.ebsco.com
Connect on LinkedIn